Reconnaissance des dialectes algériens avec des modèles de langage fine-tunés

Abstract

Sentiment analysis has experienced a significant resurgence of interest with the rise of artificial intelligence, particularly in studying opinions expressed on social media. This thesis aims to address the challenges related to the limited availability of annotated data in Algerian dialect and to improve model accuracy in sentiment detection. To this end, fine-tuning was applied to six pre-trained language models (AraBERT, CAMeLBERT, QARiB, mBERT, XLM, DistilBERT) for binary classification (positive/negative). The QARiB model achieved the best results with an accuracy of 91.1%. A cross-evaluation on Moroccan and Tunisian dialect corpora was conducted to assess the models’ generalization ability. This work makes a significant contribution to the automatic processing of the Algerian dialect and paves the way for future research on multi-dialectal models